Qwen3.5-2B-Base 昇腾适配验证报告

+------------------------------------------------------------------------------------------------+
| npu-smi 25.5.1                   Version: 25.5.1                                               |
+---------------------------+---------------+----------------------------------------------------+
| NPU   Name                | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
| Chip                      | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
+===========================+===============+====================================================+
| 3     910B4               | OK            | 91.9        40                0    / 0             |
| 0                         | 0000:02:00.0  | 0           0    / 0          17790/ 32768         |
+===========================+===============+====================================================+
+---------------------------+---------------+----------------------------------------------------+
| NPU     Chip              | Process id    | Process name             | Process memory(MB)      |
+===========================+===============+====================================================+
| 3       0                 | 2756          | VLLMEngineCor            | 14973                   |
+===========================+===============+====================================================+

vllm serve /opt/atomgit/Qwen3.5-2B-Base \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 1 \
  --max-model-len 32768 \
  --trust-remote-code \
  --gpu-memory-utilization 0.90

INFO  [vllm.py:754] Asynchronous scheduling is enabled.
INFO  [core.py:103] Initializing a V1 LLM engine (v0.18.0) with config:
      model='/opt/atomgit/Qwen3.5-2B-Base'
      dtype=torch.bfloat16
      max_seq_len=32768
      tensor_parallel_size=1
      device_config=npu
INFO  [platform.py:354] PIECEWISE compilation enabled on NPU.
      use_inductor not supported - using only ACL Graph mode
INFO  [compilation.py:289] Enabled custom fusions: norm_quant, act_quant
INFO  [backends.py:988] Using cache directory for vLLM's torch.compile
INFO  [compiler_interface.py:162] enable_npugraph_ex is enabled,
      which will bring graph compilation optimization.

{
  "object": "list",
  "data": [{
    "id": "/opt/atomgit/Qwen3.5-2B-Base",
    "object": "model",
    "owned_by": "vllm",
    "root": "/opt/atomgit/Qwen3.5-2B-Base",
    "max_model_len": 32768
  }]
}

{
  "model": "/opt/atomgit/Qwen3.5-2B-Base",
  "prompt": "The future of artificial intelligence is",
  "max_tokens": 50,
  "temperature": 0.7
}

{
  "id": "cmpl-a2fa2b603b53b8b2",
  "object": "text_completion",
  "model": "/opt/atomgit/Qwen3.5-2B-Base",
  "choices": [{
    "index": 0,
    "text": " not coming\n\nThere's a lot of speculation on the future of artificial intelligence, but these predictions are far too optimistic and, as such, can be easily dismissed.\n\nIn the 1990s, artificial intelligence was the darling of the",
    "finish_reason": "length"
  }],
  "usage": {
    "prompt_tokens": 6,
    "completion_tokens": 50,
    "total_tokens": 56
  }
}

{
  "model": "/opt/atomgit/Qwen3.5-2B-Base",
  "messages": [{"role": "user", "content": "What is the capital of France?"}],
  "max_tokens": 50,
  "temperature": 0.7
}

{
  "id": "chatcmpl-ad58bbd2c30316d9",
  "object": "chat.completion",
  "model": "/opt/atomgit/Qwen3.5-2B-Base",
  "choices": [{
    "index": 0,
    "message": {
      "role": "assistant",
      "content": "The capital of France is **Paris**.\n"
    },
    "finish_reason": "stop"
  }],
  "usage": {
    "prompt_tokens": 19,
    "completion_tokens": 11,
    "total_tokens": 30
  }
}

# 启动命令
vllm serve /opt/atomgit/Qwen3.5-2B-Base \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 1 \
  --max-model-len 32768 \
  --trust-remote-code \
  --gpu-memory-utilization 0.90

# 环境检查
$ npu-smi info
# 输出: 1× Ascend 910B4, Health OK

$ pip list | grep -E "(vllm|ascend)"
# 输出: vllm 0.18.0+empty, vllm_ascend 0.18.0rc1, triton-ascend 3.2.0.dev20260322

# 服务启动
$ vllm serve /opt/atomgit/Qwen3.5-2B-Base \
    --host 0.0.0.0 --port 8000 \
    --tensor-parallel-size 1 \
    --max-model-len 32768 \
    --trust-remote-code \
    --gpu-memory-utilization 0.90

# API 测试
$ curl http://localhost:8000/v1/models
# 输出: {"data":[{"id":"/opt/atomgit/Qwen3.5-2B-Base","max_model_len":32768}]}

$ curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "/opt/atomgit/Qwen3.5-2B-Base", "messages": [{"role": "user", "content": "What is the capital of France?"}], "max_tokens": 50}'
# 输出: {"choices":[{"message":{"content":"The capital of France is **Paris**."}}]}